research

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Research

研究

Own

RESEARCH.md

. Use it as the ground truth for project state.

负责维护

RESEARCH.md

文件，将其作为项目状态的基准依据。

Session Start

会话启动

RESEARCH.md

does not exist, initialize the project:

If
```
templates/RESEARCH.md.template
```
exists locally, copy it.

Otherwise, download it (and

CLAUDE.md

) from the pack's GitHub repo:

bash

mkdir -p templates
curl -fsSL https://raw.githubusercontent.com/zy-ning/Oh_My_Co-Researcher/main/templates/RESEARCH.md.template -o templates/RESEARCH.md.template
curl -fsSL https://raw.githubusercontent.com/zy-ning/Oh_My_Co-Researcher/main/CLAUDE.md -o CLAUDE.md
cp templates/RESEARCH.md.template RESEARCH.md

Ask the user to fill in
```
## Goal
```
, then proceed.

If
```
.co-researcher/skills.yaml
```
exists, read it before planning so you know the project's preferred skillpacks, subset choices, and supervision defaults.
Read Pipeline Status first (30-second orient), then Supervision Policy if present, then the full document.

若

RESEARCH.md

不存在，则初始化项目：

若本地存在
```
templates/RESEARCH.md.template
```
，直接复制该文件。

否则，从工具包的GitHub仓库下载

templates/RESEARCH.md.template

和

CLAUDE.md

：

bash

mkdir -p templates
curl -fsSL https://raw.githubusercontent.com/zy-ning/Oh_My_Co-Researcher/main/templates/RESEARCH.md.template -o templates/RESEARCH.md.template
curl -fsSL https://raw.githubusercontent.com/zy-ning/Oh_My_Co-Researcher/main/CLAUDE.md -o CLAUDE.md
cp templates/RESEARCH.md.template RESEARCH.md

请用户填写
```
## Goal
```
部分，之后继续执行流程。

若
```
.co-researcher/skills.yaml
```
存在，在制定计划前先读取该文件，了解项目偏好的技能包、子集选择以及监督默认设置。
先快速查看Pipeline Status（30秒内完成定位），若存在Supervision Policy则接着查看，最后完整阅读整个文档。

Supervision Policy

监督策略

```
## Supervision Policy
```
in RESEARCH.md is optional. If absent, default to confirmation-first behavior throughout unless
```
.co-researcher/skills.yaml
```
specifies project-local supervision defaults.
Precedence rule:
- ```
RESEARCH.md
```
```
## Supervision Policy
```
  overrides everything when present.
- ```
.co-researcher/skills.yaml
```
  provides project-local defaults when RESEARCH.md does not define supervision policy yet.
- preset values are only the starting recommendation written by
```
customize
```
  .
If
```
.co-researcher/skills.yaml
```
exists, treat it as the project-local preference layer for preferred presets, enabled packs, selected donor skills, and supervision defaults.
If the user asks to configure automation, supervision, or preferred skillpacks, use
```
customize
```
as the primary flow. Use direct inline RESEARCH.md edits only for lightweight policy adjustments when the user clearly wants that.
Policy dimensions:
- Notify — events to surface; if no channel exists, surface in the next user-visible update
- Approve — action types requiring a pause; precedence: approval > notify-only > silent
- Stop — target or limits (
```
target_reached
```
  ,
```
step_limit
```
  ,
```
time_budget
```
  ,
```
error_threshold
```
  ,
```
blocked
```
  )
- Resources — three classes: Service/API, Compute, Human/Physical; escalation rule is
```
forbid
```
  /
```
notify
```
  /
```
approve
```
- Idea Changes — improvements (notify), strategy pivots (notify or approve), compromises (approve by default)
Preset behavior:
- ```
manual
```
  — always stop for approval
- ```
checkpointed
```
  (default) — record approval-required issues in Pending Approvals, continue unrelated allowed work; stop only when nothing else remains
- ```
autonomous
```
  — notify but do not stop
- ```
wild
```
  — never implicit; warn the user about unattended actions, resource use, and plan divergence, then require explicit confirmation before enabling

```
RESEARCH.md
```
中的
```
## Supervision Policy
```
为可选内容。若未设置，则全程默认采用“先确认再执行”的行为模式，除非
```
.co-researcher/skills.yaml
```
中指定了项目本地的监督默认设置。
优先级规则：
- 若
```
RESEARCH.md
```
  中存在
```
## Supervision Policy
```
  ，则其优先级最高，覆盖所有其他设置。
- 当
```
RESEARCH.md
```
  尚未定义监督策略时，
```
.co-researcher/skills.yaml
```
  提供项目本地的默认设置。
- 预设值仅为
```
customize
```
  功能写入的初始推荐配置。
若
```
.co-researcher/skills.yaml
```
存在，则将其视为项目本地的偏好层，用于设置首选预设、启用的工具包、选定的捐赠技能以及监督默认值。
若用户要求配置自动化、监督或偏好技能包，将
```
customize
```
作为主要流程。仅当用户明确要求时，才直接对
```
RESEARCH.md
```
进行轻量级的策略调整编辑。
策略维度：
- Notify（通知） — 需要主动告知的事件；若没有指定通知渠道，则在下次用户可见的更新中展示相关内容
- Approve（审批） — 需要暂停等待审批的操作类型；优先级：审批 > 仅通知 > 静默执行
- Stop（停止） — 触发停止的目标或限制条件（
```
target_reached
```
  、
```
step_limit
```
  、
```
time_budget
```
  、
```
error_threshold
```
  、
```
blocked
```
  ）
- Resources（资源） — 三类资源：服务/API、计算资源、人力/物理资源；升级规则为
```
禁止
```
  /
```
通知
```
  /
```
审批
```
- Idea Changes（想法变更） — 改进建议（通知）、策略转向（通知或审批）、折中方案（默认需审批）
预设行为模式：
- ```
manual
```
  — 始终停止操作等待审批
- ```
checkpointed
```
  （默认） — 将需要审批的问题记录在Pending Approvals中，继续执行其他允许的工作；仅当没有其他可执行任务时才停止
- ```
autonomous
```
  — 发送通知但不停止操作
- ```
wild
```
  — 绝不自动执行；先警告用户无人值守操作、资源使用情况以及计划偏离问题，然后在启用前需要用户明确确认

Assess Before Acting

执行前评估

Before executing any TODO, do a gap analysis:

Compare Goal against completed Context entries. What's missing?
Surface stale Blocked items immediately.
Gaps to catch: experiment done but no result analysis; results done but no paper; paper drafted but no review.

If the TODO list is stale or incomplete, apply the supervision policy for
```
todo-adjustments
```
(approve → wait; notify → announce and continue; absent → propose and wait), then add/remove/reorder as needed.

在执行任何TODO任务前，进行缺口分析：
- 对比Goal与已完成的Context条目，找出缺失内容。
- 立即标记过期的Blocked条目。
- 需要排查的缺口：已完成实验但未分析结果；已得到结果但未撰写论文；已完成论文草稿但未进行评审。
若TODO列表过期或不完整，针对
```
todo-adjustments
```
应用监督策略（审批→等待；通知→告知并继续；未设置→提出建议并等待），之后按需添加/删除/重新排序任务。

Act

执行任务

Pick the highest-priority unblocked TODO. Announce:
```
→ Working on: <TODO text>
```
.
Delegate to the right skill:

Need	Skill
Literature survey	`research-lit` or `arxiv`
Deep paper Q&A / code repo discovery	`alpha-research` (AlphaXiv: `alpha ask` , `alpha repo` )
Paper-to-code reproducibility audit	`audit`
Method / idea refinement	`research-refine`
Experiment planning	`experiment-plan`
Running experiments	`experiment` (this pack) or `run-experiment`
Result analysis	`result-to-claim`
Figures	`paper-figure`
Paper drafting	`write` (this pack) or `paper-write`
Review / critique	`review` (this pack) or `auto-review-loop`
Lesson extraction	`evolve` (this pack)
Unclear	Ask the user, or do it directly.

When done: check off TODO, append timestamped Context entry, update State and Pipeline Status, save
```
RESEARCH.md
```
.

选择优先级最高且未被阻塞的TODO任务，告知用户：
```
→ 正在处理：<TODO内容>
```
。

将任务委派给合适的技能：

需求	技能
文献调研	`research-lit` 或 `arxiv`
深度论文问答/代码仓库发现	`alpha-research` （AlphaXiv： `alpha ask` 、 `alpha repo` ）
论文转代码可复现性审核	`audit`
方法/想法优化	`research-refine`
实验规划	`experiment-plan`
运行实验	`experiment` （本工具包）或 `run-experiment`
结果分析	`result-to-claim`
图表制作	`paper-figure`
论文撰写	`write` （本工具包）或 `paper-write`
评审/批评	`review` （本工具包）或 `auto-review-loop`
经验提取	`evolve` （本工具包）
需求不明确	询问用户，或直接执行。

任务完成后：勾选TODO任务，添加带时间戳的Context条目，更新State和Pipeline Status，保存
```
RESEARCH.md
```
。

Proactive Loop

主动循环

After each completed TODO, surface the result briefly, then apply the supervision policy for the next move:

Ambiguous next step → describe 2–3 options and ask. Never silently pick among plausible directions.
Approve / stop target reached → ask before proceeding, or queue and continue unrelated work if
```
checkpointed
```
.
Notify only → announce the next action and continue.
Wild → continue until completion criteria, a hard safety rule, or a forbidden boundary.
No policy → announce and ask "Shall I continue?"

每个TODO任务完成后，简要展示结果，然后针对下一步操作应用监督策略：
- 下一步操作不明确 → 描述2-3种选项并询问用户。绝不在合理的方向中自行选择。
- 达到审批/停止目标 → 继续执行前询问用户；若为
```
checkpointed
```
  模式，则将任务加入队列并继续执行其他无关工作。
- 仅通知 → 告知下一步操作并继续执行。
- Wild模式 → 持续执行直至满足完成标准、触发硬性安全规则或触及禁止边界。
- 无策略设置 → 告知用户并询问“是否继续？”

Exploration Modes

探索模式

BFS mode — opt-in, off by default

When the user asks to "explore the design space", "autoresearch", or "find the best config", apply the supervision policy for

bfs-start

(approve → confirm; notify → announce and continue; absent → confirm). In

wild

mode, BFS may proceed without a checkpoint only when target file, metric, budget, and completion criteria are already explicit and within granted resource boundaries.

Before starting, confirm: target file, metric + direction (e.g., minimize

val_bpb

), time budget per run, max experiments. Then delegate to

experiment

BFS mode — do not interrupt. When done, update RESEARCH.md Context and Pipeline Status and propose next step.

DFS — default when direction is confirmed

Commit each meaningful step.

git stash

before risky moves; note the stash in Context.

BFS模式 — 可选启用，默认关闭

当用户要求“explore the design space”、“autoresearch”或“find the best config”时，针对

bfs-start

应用监督策略（审批→确认；通知→告知并继续；未设置→确认）。在

wild

模式下，仅当目标文件、指标、预算和完成标准已明确且在允许的资源范围内时，BFS模式才可无需检查点直接执行。

启动前确认：目标文件、指标+方向（例如，最小化

val_bpb

）、每次运行的时间预算、最大实验次数。然后将任务委派给

experiment

的BFS模式 — 请勿中断。完成后，更新

RESEARCH.md

的Context和Pipeline Status，并提出下一步建议。

DFS — 方向明确时的默认模式

执行每一个有意义的步骤。在执行高风险操作前执行

git stash

；在Context中记录stash信息。

Project-local Skill Config

项目本地技能配置

```
.co-researcher/skills.yaml
```
is optional. If absent, keep existing behavior.
If present, use it to prefer enabled donor packs and selected skills when choosing among overlapping external tools.
Do not silently rewrite
```
.co-researcher/skills.yaml
```
during normal orchestration. Use
```
customize
```
when the user wants to change stack or supervision preferences.

```
.co-researcher/skills.yaml
```
为可选文件。若不存在，则保持现有行为模式。
若存在，在选择重叠的外部工具时，优先使用其中启用的捐赠工具包和选定的技能。
在正常编排过程中，请勿擅自改写
```
.co-researcher/skills.yaml
```
。当用户想要更改技术栈或监督偏好时，使用
```
customize
```
功能。

Tone

语气

Concise. No preamble. Show what you are doing, not what you are about to do.

简洁直接，无需开场白。展示正在执行的操作，而非即将执行的操作。

Example

示例

Input: RESEARCH.md with experiments done in Context, but only a "run baseline" TODO (already checked). No write TODO. Gap detected: "Results exist but no paper TODO — adding P1: Write paper draft." Action: Proposes TODO addition, waits for confirmation, then invokes

result-to-claim

to validate claims before writing.

输入：

RESEARCH.md

的Context中记录已完成实验，但TODO列表只有一个已勾选的“run baseline”任务，没有撰写论文的TODO。检测到缺口：“已存在实验结果但无撰写论文的TODO — 添加优先级1任务：撰写论文草稿。” 操作：提出添加TODO的建议，等待用户确认，然后调用

result-to-claim

在撰写前验证结论。